infosave/MOSS-Image-SFT-cmf

MOSS-Image-SFT CMF

Native, single-file CMF packaging of OpenMOSS-Team/MOSS-Image-SFT at source revision f972bd4b4b9e3852133cc5da954d87bc32cd8fb9, for the Rust cortiq runtime.

MOSS-Image-SFT is the supervised fine-tuned text-to-image model of the MOSS-Image family, derived from MOSS-Image-Base with the same architecture and inference pipeline: natural-language prompts, tag sequences and JSON layout prompts. This repository is the standalone CMF artifact for the SFT weights; it does not change the cards or artifacts of MOSS-Image-Base or any other model. Native MOSS text-to-image support requires Cortiq v0.8.31 or newer.

English · 中文

Examples

Four full 50-step q8_2f renders on the named Vulkan reference system (RTX 2000 Ada, driver 595.91.07), visually inspected before publication. The command under each image reproduces it exactly, with its PNG SHA-256.

Red fox by a forest stream

PROMPT='best quality, highres, RAW editorial wildlife photograph of an adult red fox standing on a moss-covered stone beside a clear forest stream at dawn, detailed natural fur, amber eyes, soft mist, autumn ferns, golden backlight, 85mm lens, shallow depth of field, realistic cinematic color grading'

cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "$PROMPT" \
  --width 512 --height 512 --steps 50 --cfg 3 --shift 5 \
  --seed 20261024 --out samples/fox-sft-512-q8-50s.png

PNG SHA-256: 09bb563198602439b0bec415fdb37da08bb37e4d0f6d4b1b1ebe629b482c0c5a

Rainy neon alley

PROMPT='best quality, highres, cinematic photograph of a rainy neon-lit Tokyo alley at night, wet asphalt reflections, red and blue neon signs, a single umbrella-carrying pedestrian seen from behind, 35mm lens, deep contrast, film grain'

cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "$PROMPT" \
  --width 512 --height 512 --steps 50 --cfg 3 --shift 5 \
  --seed 20261025 --out samples/tokyo-alley-sft-512-q8-50s.png

PNG SHA-256: 095d5f36edb2af5a86e9a871c18a69110ee8034552ca61fa3da80a3548f389d6

Still life with peaches

PROMPT='best quality, highres, studio still life photograph of ripe peaches and a glass of water on a linen cloth, soft window light from the left, delicate shadows, shallow depth of field, warm color palette, 85mm macro'

cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "$PROMPT" \
  --width 512 --height 512 --steps 50 --cfg 3 --shift 5 \
  --seed 20261026 --out samples/stilllife-sft-512-q8-50s.png

PNG SHA-256: ed9fd65e42ded1d1637dba06ed1c517766ec494dec8894c759e0d00614542f0b

Red squirrel in an autumn park

PROMPT='best quality, highres, RAW wildlife photograph of a red squirrel sitting on an old oak log in an autumn park, holding a nut in its paws, bushy tail curled up, golden fallen leaves around, warm morning sunlight, 200mm lens, shallow depth of field, detailed fur texture'

cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "$PROMPT" \
  --width 512 --height 512 --steps 50 --cfg 3 --shift 5 \
  --seed 20261027 --out samples/squirrel-sft-512-q8-50s.png

PNG SHA-256: 9d42ffd1ac81b1c3a99b8ec472975e05149dc0dc68651cae40c6fcd001d411af

512 × 512 is the source-supported lower resolution boundary; use the 1–2 MP guidance in Recommended quality recipe when the hardware budget permits it.

Downloads and integrity

Artifact	Size	SHA-256	Use
moss-image-sft-q8_2f.cmf	13,256,239,565 bytes (12.35 GiB)	f63408c573ac2d146da8aea324253983d0d83d739662b44765105383a5234439	Recommended, quality-first profile
moss-image-sft-q4tp.cmf	10,856,278,477 bytes (10.11 GiB)	22d858e809f1c030f87496abd9dd8799af00de9e37776640235870f318bd7490	Lower DiT-memory-footprint profile

A matching .sha256 file is published beside each CMF. Verify a download before use:

REPO=infosave/MOSS-Image-SFT-cmf
MODEL=moss-image-sft-q8_2f.cmf       # or moss-image-sft-q4tp.cmf

hf download "$REPO" "$MODEL" "$MODEL.sha256" --local-dir .
sha256sum -c "$MODEL.sha256"          # macOS: shasum -a 256 -c "$MODEL.sha256"
cortiq verify "$MODEL"

What the CMF runs

The packaged SFT text-to-image path is:

1. the dense hybrid Qwen3.5 text encoder with its tokenizer and stored chat template;
2. the 42-layer MOSSImageTransformer (width 3072, 24 heads, FFN width 8192); and
3. the FLUX.2 AutoencoderKLFlux2 decode path.

Both files contain 1038 tensors (463 DiT, 426 text-encoder, 142 Flux2 decode). The tokenizer, template, scheduler, and component configuration are embedded. The runtime uses staged residency — text encoder, then DiT, then decoder — so the previous stage can be released before the next is resident.

As on the Base artifact, the CMF packs only the components reached by the observed string-prompt text-to-image route. The Qwen visual tower and the encoder-side Flux2/VAE path (including encoder-side quant_conv weights) are not included because this route does not call them. This is a scope decision, not a claim that those source components are unnecessary in every upstream workflow. Image-conditioned, editing, video and reference-image workflows are not advertised by this CMF.

Prompting and JSON layouts

MOSS-Image-SFT supports the upstream prompt forms:

• ordinary natural-language descriptions;
• tag sequences; and
• a raw JSON layout prompt for text placement and object arrangement.

Use natural language or tags with --prompt. With a build whose cortiq imagine --help exposes --layout / --layout-file (v0.8.31+), use exactly one of them instead. The CLI validates that the input is a JSON object and passes its text unchanged through the stored official MOSS prompt template. The upstream layout convention documents summary.caption, items, bbox_format: "yxyx" and bbox_coord_space: "abs_centralized"; use the upstream schema and matching output dimensions.

cortiq imagine moss-image-sft-q8_2f.cmf \
  --layout-file layout.json \
  --width 1024 --height 1024 --out layout.png

Layout editing means editing the JSON caption, items or bounding boxes and then generating a new image. It is not pixel-space editing or img2img. In particular, this CMF has no MOSS --image input.

Recommended quality recipe

The embedded source recipe is 1024 × 1024, 50 Flow-Matching Euler steps, CFG 3, scheduler shift 5, an empty negative prompt, and the source negative-branch skip at zero-based DiT layer 22. Width and height must be positive multiples of 16. This matches the upstream SFT quick start (guidance_scale=3.0, uncond_skip_layers=[22]).

For quality output, follow the upstream guidance: use at least 50 steps and about 1–2 million pixels. The default 1024 × 1024 setting is about 1.05 million pixels. Faster, lower-resolution settings are useful for functional smokes but are not a substitute for this quality recipe.

cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "A small red fox reading a book in a sunlit library, highly detailed." \
  --width 1024 --height 1024 --steps 50 --cfg 3 --shift 5 \
  --out fox.png

Use q8_2f when output fidelity is the priority. Choose q4tp when a smaller DiT footprint is more important, and evaluate it on the intended prompts rather than assuming it is visually interchangeable with q8_2f.

Precision profiles

Profile	DiT matrices	Text conditioning	Decoder and control tensors
q8_2f	q8_2f	q8_2f projections; q8-row embeddings	Flux2 decoder, norms, biases, and time/modulation controls remain floating-point
q4tp	q4tp	q8_2f projections; q8-row embeddings	Same floating-point decoder and controls

The two published files differ only in the large DiT matrix codec; the text-conditioning and decoder/control choices stay identical. For an explicit, inspectable compromise, imagine-pack --dit-keep can lift selected MOSS DiT blocks or projections to q8_2f over a q4tp base:

cortiq imagine-pack /path/to/MOSS-Image-SFT \
  --quant q4tp --dit-keep layers.0,layers.1,layers.40,layers.41 \
  --out moss-image-sft-q4tp-guarded.cmf

q4tp-v2 is intentionally not used as a filename here: in this project that name is reserved for a data-aware, calibrated/refit q4 conversion, and no such image-DiT calibration has been measured for SFT.

MOSS controls in Cortiq v0.8.31+

Confirm the installed binary with cortiq imagine --help.

• --uncond-skip-layers LAYERS selects zero-based negative-branch layers to skip during the active interval. The packaged source default is layer 22. A bare flag, none or off disables that default.
• --slg-layers LAYERS and --slg-scale SCALE enable standard skip-layer-guidance (SLG) correction on selected positive-branch layers. When active, SLG adds a third DiT forward, so it increases render time.
• --slg-start F and --slg-stop F set the active denoising-fraction interval from 0 to 1 for the source skip/SLG behavior.

# Disable the packaged negative-branch skip.
cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "A lacquered red teapot on a wooden table" \
  --uncond-skip-layers none --out no-skip.png

# Example SLG configuration; choose layer indices and scale for the target prompt.
cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "A lacquered red teapot on a wooden table" \
  --slg-layers 24,26 --slg-scale 3 --slg-start 0 --slg-stop 1 \
  --out slg.png

Validation and measured baselines

Both files passed cortiq verify (1038 tensors, all hashes match) and an external SHA-256 check after the pack, and both were exercised end-to-end on the Vulkan reference system: a functional 2-step smoke (output inspected, not published as a quality example) and the 8-step CFG baseline below.

Vulkan baseline

Measured on an NVIDIA RTX 2000 Ada (16 GB), Vulkan driver 595.91.07, Cortiq v0.8.31: 512 × 512, 8 steps, CFG 3, shift 5, seed 42. median DiT step includes the two DiT passes required by CFG. Text preparation, DiT loading and VAE decoding are reported separately by the CLI.

File	Total	Median DiT step	Text / DiT load / VAE
q8_2f	276.2 s	32.42 s	2.3 s / 0.2 s / 11.9 s
q4tp	266.3 s	31.21 s	2.3 s / 0.2 s / 11.9 s

This is a named, reproducible baseline, not a general performance guarantee for other drivers, hardware, resolution, precision profiles or settings. The GPU path reported weight budget 13820 MB, so both files fit a 16 GB card with room for a 1024 × 1024 context.

Compatibility and scope

This publication supports the SFT text-to-image path, including natural-language prompts, tag sequences and serialized JSON-layout prompts (v0.8.31+ builds exposing the layout flags). Its scope ends there. It does not claim support for:

• image input, img2img, or pixel-space image editing;
• video or reference-image workflows;
• LoRA adapter loading or training;
• the upstream Gradio application, ComfyUI integration, prompt-enhancer UI, CPU-offload/compilation options, or other Python-only surfaces; or
• bit-exact numerical parity with the upstream PyTorch implementation.

Quick start

cargo install cortiq-cli
hf download infosave/MOSS-Image-SFT-cmf moss-image-sft-q8_2f.cmf --local-dir .
cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "A small red fox reading a book in a sunlit library, highly detailed." \
  --width 1024 --height 1024 --steps 50 --cfg 3 --shift 5 --out fox.png

License and provenance

The upstream MOSS-Image-SFT release is licensed under Apache-2.0. These CMF artifacts are quantized/repackaged derivatives and retain upstream license and attribution. Consult the upstream repository and model card for model behavior, dataset disclosures and any use restrictions.

---

中文说明

OpenMOSS-Team/MOSS-Image-SFT （源版本 f972bd4b4b9e3852133cc5da954d87bc32cd8fb9）打包为独立的 CMF 单文件，由 Rust 推理引擎 cortiq 运行。MOSS-Image-SFT 是 MOSS-Image 家族中由 MOSS-Image-Base 监督微调得到的文生图模型，架构与推理管线完全相同：自然语言提示、标签序列 与 JSON 布局提示。需要 Cortiq v0.8.31 或更新版本。

下载与校验

文件	大小	SHA-256	用途
moss-image-sft-q8_2f.cmf	13,256,239,565 字节（12.35 GiB）	f63408c573ac2d146da8aea324253983d0d83d739662b44765105383a5234439	推荐，画质优先
moss-image-sft-q4tp.cmf	10,856,278,477 字节（10.11 GiB）	22d858e809f1c030f87496abd9dd8799af00de9e37776640235870f318bd7490	更小的 DiT 显存占用

hf download infosave/MOSS-Image-SFT-cmf moss-image-sft-q8_2f.cmf --local-dir .
sha256sum -c moss-image-sft-q8_2f.cmf.sha256
cortiq verify moss-image-sft-q8_2f.cmf

内容与配方

CMF 内含：Qwen3.5 文本编码器、42 层 MOSSImageTransformer（宽 3072、24 头、 FFN 8192）、FLUX.2 AutoencoderKLFlux2 解码路径，共 1038 个张量（DiT 463、 文本编码器 426、解码器 142）。源配方：1024×1024、50 步、CFG 3、shift 5、 负分支跳过第 22 层——与上游 SFT 快速开始一致。仅打包字符串提示文生图路径 可达的组件；不支持图像输入、img2img、视频、LoRA 与位级数值一致。

示例（均为 50 步 q8_2f 渲染，RTX 2000 Ada / Vulkan）

基线

RTX 2000 Ada（16 GB）、Vulkan 驱动 595.91.07、Cortiq v0.8.31、512×512、 8 步、CFG 3、shift 5、seed 42：

文件	总计	DiT 步中位数	文本 / DiT 加载 / VAE
q8_2f	276.2 秒	32.42 秒	2.3 秒 / 0.2 秒 / 11.9 秒
q4tp	266.3 秒	31.21 秒	2.3 秒 / 0.2 秒 / 11.9 秒

快速开始

cargo install cortiq-cli
hf download infosave/MOSS-Image-SFT-cmf moss-image-sft-q8_2f.cmf --local-dir .
cortiq imagine moss-image-sft-q8_2f.cmf \
  --prompt "一只在阳光书房里读书的小红狐，细节丰富。" \
  --width 1024 --height 1024 --steps 50 --cfg 3 --shift 5 --out fox.png

许可与来源

上游 MOSS-Image-SFT 以 Apache-2.0 发布；本 CMF 为其量化再打包衍生版本，保留上游许可与署名。